Live Feed/vLLM/Fact Record
vLLM logo
vLLM
feature 96% Confidence Gate October 17, 2026

v0.32.0rc3: [Bugfix] Fix NaN output of the Mamba2 chunk scan kernel on sm90 (#61718)

This release resolves a critical numerical stability issue causing NaN outputs in the Mamba2 chunk scan kernel. The fix specifically targets hardware architectures utilizing the sm90 compute capability.

Verified State Diff

Comparison Mode:
- Previous State
Mamba2 chunk scan kernels produced NaN outputs when running on sm90 (NVIDIA Hopper) hardware.
+ Verified New State
Mamba2 chunk scan kernels produce numerically stable outputs on sm90 hardware.

Impact & Verification Analysis

WHO IS AFFECTED

Developers and researchers deploying Mamba2 models on NVIDIA Hopper-based GPUs.

WHY IT MATTERS

Ensures functional parity and reliability for state-space models on high-performance hardware, preventing silent failures in production inference pipelines.

Full Fact Overview

The update addresses a bug in the Mamba2 implementation within vLLM, specifically affecting the chunk scan kernel when executed on NVIDIA Hopper (sm90) architecture. By correcting the underlying kernel logic, the patch ensures that state-space model (SSM) computations maintain numerical integrity during inference, preventing the generation of Not-a-Number (NaN) values that previously rendered model outputs unusable on this hardware.

Multi-Source Evidence Chain (1)

v0.32.0rc3: [Bugfix] Fix NaN output of the Mamba2 chunk scan kernel on sm90 (#61718)vLLM
TRACKED ENTITY
Explore all historical vLLM changes
View vLLM Hub ➔